Zhendong Liu

We propose VoTA, a novel attack framework that exploits the tension between logical reasoning and safety objectives in VLMs by generating chains of images with risky visual thoughts, achieving significantly higher success rates than existing methods.